Download Remote sensing of water depths in shallow - Everglades-HUB

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Abyssal plain wikipedia , lookup

Global Energy and Water Cycle Experiment wikipedia , lookup

Critical Depth wikipedia , lookup

Transcript
Estuarine, Coastal and Shelf Science 89 (2010) 89e96
Contents lists available at ScienceDirect
Estuarine, Coastal and Shelf Science
journal homepage: www.elsevier.com/locate/ecss
Remote sensing of water depths in shallow waters via artificial neural networks
Özçelik Ceyhun*, Arısoy Yalçın
Department of Civil Engineering, Dokuz Eylul University, 35160 Izmir, Turkey
a r t i c l e i n f o
a b s t r a c t
Article history:
Received 25 November 2009
Accepted 26 May 2010
Available online 2 June 2010
Determination of the water depths in coastal zones is a common requirement for the majority of coastal
engineering and coastal science applications. However, production of high quality bathymetric maps
requires expensive field survey, high technology equipment and expert personnel. Remotely sensed
images can be conveniently used to reduce the cost and labor needed for bathymetric measurements and
to overcome the difficulties in spatial and temporal depth provision. An Artificial Neural Network (ANN)
methodology is introduced in this study to derive bathymetric maps in shallow waters via remote
sensing images and sample depth measurements. This methodology provides fast and practical solution
for depth estimation in shallow waters, coupling temporal and spatial capabilities of remote sensing
imagery with modeling flexibility of ANN. Its main advantage in practice is that it enables to directly use
image reflectance values in depth estimations, without refining depth-caused scatterings from other
environmental factors (e.g. bottom material and vegetation). Its function-free structure allows evaluating
nonlinear relationships between multi-band images and in-situ depth measurements, therefore leads
more reliable depth estimations than classical regressive approaches. The west coast of the Foca, Izmir/
Turkey was used as a test bed. Aster first three band images and Quickbird pan-sharpened images were
used to derive ANN based bathymetric maps of this study area. In-situ depth measurements were
supplied from the General Command of Mapping, Turkey (HGK). Two models were set, one for Aster and
one for Quickbird image inputs. Bathymetric maps relying solely on in-situ depth measurements were
used to evaluate resultant derived bathymetric maps. The efficiency of the methodology was discussed at
the end of the paper. It is concluded that the proposed methodology could decrease spatial and repetitive
depth measurement requirements in bathymetric mapping especially for preliminary engineering
application.
Ó 2010 Elsevier Ltd. All rights reserved.
Keywords:
water depth
bathymetry
remote sensing
artificial neural networks
1. Introduction
Bathymetric surveying of shallow waters has a great importance
for coastal engineering and coastal science applications as well as
shipping safety (Leu and Chang, 2005; Grilli, 1998). Especially in
coastal zones intensive sediment transportation due to tidal
movements, wave propagation, bottom currents, tributaries etc.
cause significant temporal and spatial changes on sea bottom and
made recursive surveying necessary (Zanial, 1994; Spiter and Dirks,
1987; Lyzenga, 1985, 1978). Therefore, a requirement for a practical
depth estimation method arises especially for preliminary
applications.
The shipboard echo sounder having been used conventionally to
measure water depths gives quite accurate results for point
measurements (Leu and Chang, 2005; Lyzenga, 1985). However,
* Corresponding author.
E-mail address: [email protected] (Ö. Ceyhun).
0272-7714/$ e see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ecss.2010.05.015
since it needs intensive labor, time and money, remote sensing
techniques including airborne lidar measurements and optical
remote sensing are preferably used in practical coastal engineering
and coastal science applications. (Mas, 2004; Martin, 1993; Lyon
et al., 1992). Lidar has a high accuracy in depth measurements,
provided that the altitude of the measurement platform is known
accurately. Its coverage is limited by maximum altitude, scan angle
and position of the platform (Lyzenga, 1985, 1981). Optical remote
sensing relies on passive multispectral scanner data. It uses optical
characteristics of the water column to estimate water depths
(Fonstad and Marcus, 2005; Lyzenga, 1978). Since multispectral
scanner images can be easily obtained for different time and location, it is favored for practical bathymetric data requirements
(Louchard et al., 2003).
Various depth estimation methods based on the optical remote
sensing developed in the literature. For instance, Lyzenga (1978)
proposed a modified exponential depth model for clear shallow
waters, ignoring the internal reflection in the water column;
Louchard et al. (2003) realized radiative transfer calculations to
90
Ö. Ceyhun, A. Yalçın / Estuarine, Coastal and Shelf Science 89 (2010) 89e96
generate a spectral library of remote sensing reflectance and thus to
classify obtained reflectance according to bottom type and water
depth; Leu and Chang (2005) used two dimensional wave spectrums to estimate water depths based on the principle that while
waves propagate toward shoreline, their wave lengths decrease due
to decrements in water depth; Fonstad and Marcus (2005)
combined remote sensing imagery and open channel flow principals to estimate water depths in clear rivers. Most of these methods
are either difficult to apply or valid for only some specific conditions. Therefore, linear regression models on spectral reflectance
values are generally more attractive for practical purposes because
they are easy to handle as well as they can be calibrated for any area
and time images.
The Single Band Algorithm (SBA) is the simplest and most widely
used linear regression approach (Martin, 1993). The SBA assumes
that log transformed reflectance of the pixels on a single band image
are linearly correlated with water depths on those pixels. Principal
Component Algorithm (PCA) or Multi-band Approach is another
method that evaluates the relationship between in-situ depth
measurements and log transformed reflectance values. It can evaluate multi-band images together to get better depth estimations.
The PCA firstly finds the principal components (PC) of the log
transformed reflectance, and then assumes that the scores of the PC
are linearly correlated with water depths (Martin, 1993). Although
these models provide temporal and spatial freedom and easiness in
modeling, it is generally not possible to have significant linear
relationships between PC scores (or single band image reflectance)
and water depths, especially for the regions where environmental
parameters affecting the reflectance from water column change
unevenly (Leu and Chang, 2005; Fonstad and Marcus, 2005;
Lyzenga, 1981). It is very complex to refine the reflectance
sourced solely by depth changes from the effects of environmental
factors such as bottom material, vegetation, pollution, algae cover,
turbulence, tributary etc. (Hengel and Spitzer, 1991; Lyzenga, 1985).
Few researchers developed different methods to able to model
effects of these environmental factors (Härmä et al., 2001; Zanial,
1994; Lyon et al., 1992; Spiter and Dirks, 1987), however, there is
no unique reliable method that is simple and practically applied. On
the other hand, black box models, that don’t need to focus on
interior processes, stands as a reasonable alternative of regressive
models for depth estimations via optical remote sensing.
The methodology proposed in this study use Artificial Neural
Networks (ANN) to estimate water depths in shallow waters. ANN’s
nonlinear, sample based and model free structure allows the
methodology to consider nonlinear multi-parameter relationship
between reflectance from different spectral bands and water
depths. Its main advantage in practice is that it can provide depth
estimations without eliminating problems in optical remote
sensing associated with environmental factors such as bottom
material and vegetation. The proposed methodology reduces
spatial and repetitive depth measurement requirement. Technical,
logistical and economical difficulties on field surveys and on classical optical remote sensing make it useful especially for preliminary engineering applications. The Foca bay/Turkey was used as
a test bed. Two ANN models were built. Each of these models was
calibrated (trained) using Aster and Quickbird images respectively.
Bathymetric maps were derived based on these models’ estimations. Accuracies of results were evaluated based on two reference
maps relying on different number of in-situ depth measurements.
Finally, the effects of the number of the depth measurements and
different band satellite images on model estimations were
investigated.
2. Study area and data
The proposed ANN methodology is applied for the Foca bay,
Izmir, Turkey (Fig. 1). The main economic activities in the region are
fishing and tourism. Navigation is realized via the Foca port. Due to
Fig. 1. The study area, Foca district.
Ö. Ceyhun, A. Yalçın / Estuarine, Coastal and Shelf Science 89 (2010) 89e96
heavy sediment transport throughout Foca shoreline, sea bottom
have exposed to level changes, and thus navigation and other
coastal activities such as fishing and recreation have affected
negatively.
In-situ depth measurements were supplied from the surveys of
the General Command of Mapping, Turkey (HGK). Aster first three
band images and also Quickbird pan-sharpened first four band
images were used to build depth models. Both depth measurements and satellite images were acquired in same period JuneeJuly
2005. Two reference maps were generated based on all (300)/half
(150) of the depth measurements supplied, in order to see effects of
the number of the depth measurements on bathymetric mapping.
The distance-weighted average method was used for bathymetric
map generation. The resultant maps are provided in 5 m precision
to facilitate interpretations (Fig. 2).
between the [th node of the (i 1) th layer and jth node of the ith
layer; bj are the biases; Ij(p) is jth input vector (reflectance from jth
band); T(p) is the calculated output vector for each epoch; O(p) is
the expected output vector (in-situ depth measurements). For this
model, net inflow netj to the jth node of the ith layer is defined as
follows (Hagan and Menhaj, 1994; Freeman and Skapura, 1991).
netj ¼
m X
w[; j o[ þ bj
(1)
[¼1
where o[ is the output of the [th node of the (i-1) th layer and
equals to pth element of the input vectors for the first layer and to f
(netj) for the other layers. f(netj) is a transfer function that transforms the net input netj into the node outputs o[ (Hagan et al.,
1996). It is taken here as the log sigmoid function given in Eq. (2).
f netj ¼ 1= 1 þ exp netj
3. Methodology
91
(2)
Model weights are calculated by using the typical performance
function in Eq. (3) (Atkinson and Tatnall, 1997; Hagan and Menhaj,
1994).
3.1. ANN modeling
ANN having a flexible information transferring structure has
attracted large interest during the last years (Mas, 2004; Atkinson
and Tatnall, 1997; Jain et al., 1996; Lek et al., 1996). The ability to
handle nonlinear functions, estimate for unseen inputs and use
observed data have made it popular (Lek et al., 1996; Civco, 1993;
Freeman and Skapura, 1991). It has found wide range of application especially in environmental science, image processing,
medicine and molecular biology (Mas, 2004; Atkinson and Tatnall,
1997).
Connection pattern of an artificial neural network is generally
considered in two categories, feed forward and feedback (recurrent) networks (Hagan et al., 1996; Haykin, 1994). In this study
multi-layer feed forward neural network (MFN) is used for depth
estimations because of its simplicity in practice. In-situ depth
measurements are considered as the expected output vector, and
the reflectance values on each spectral band are accounted as
inputs vectors. Fig. 3 shows the general structure of the ANN model
proposed. Where i, j and p are, respectively, layer, node and
observation numbers; n, m and N are, respectively, the numbers of
layers, nodes and observations; w[;j are the network weights
RMSE ¼
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN
2
p ¼ 1 ðTðpÞ OðpÞÞ
N
(3)
If the root of mean square errors (i.e. RMSE) between calculated
and observed outputs is low enough to ignore, then initial (or
calculated) weights and biases are assumed as the resultant
weights and biases, else the calculated error is distributed via
a training algorithm to find out new weights and biases. This
procedure is repeated until an acceptable RMSE is reached. LevenbergeMarquard training algorithm given in Eq. (4) is used in this
study (Hagan and Menhaj, 1994).
h
i1
xkþ1 ¼ xk JT J þ mI
JT 3k
(4)
where xk is a vector of current weights and biases; 3 and J are,
respectively, the vector and Jacobean matrix of the network errors;
m is a scalar which indicates how to use the memory and calculation
speed of the Jacobean matrix; k is iteration number; I is the unit
matrix, and superscript T indicates transposition.
Fig. 2. Bathymetric maps generated using (a) 150, (b) 300 in-situ depth measurements.
92
Ö. Ceyhun, A. Yalçın / Estuarine, Coastal and Shelf Science 89 (2010) 89e96
Fig. 3. Typical ANN architecture used in depth estimations.
3.2. Depth modeling and simulation
Preparation of model inputs and determination of model
architecture and parameters are two important steps to model and
simulate water depths in shallow waters via the ANN model
proposed in Fig. 3. Model inputs, reflectance from different spectral
bands and in-situ measurements, should be processed before used
in the modeling. Filtering, rectifying of the input images and
scaling/rescaling of the model inputs and outputs can be regarded
in this manner. Model architecture is constrained by some factors
such as problem to be solved, number of input factors and number
of outputs required (Hagan et al., 1996). The number of the nodes in
the input layer equals to the number of input spectral bands, and
the number of the nodes in the output layer is naturally one since
there is only one output (i.e. water depths). The number of the inner
layers and nodes are determined experimentally (Freeman and
Skapura, 1991) to have the best performance model. Model
parameters are obtained as a result of training and testing
processes. Following the preparation of model inputs and determination of model architecture and parameters, water depth
simulations are made for multispectral reflectance values of
required coastal points. The methodology proposed is summarized
in Fig. 4.
According to Fig. 4, each band image should be first filtered to
both discard the land area, clouds and ships, and decrease noise and
scattering effects on input images. The land area, clouds and ships
can be discarded by using a boolean filter (referred also as a binary
or logical filter) while noises and scattering effects can be reduced
by applying pixel based filters such as mean (low pass), Gaussian,
minimum, median filters, etc. (Jain, 1989). After in-situ depth
measurements and input images are transformed into same coordinate system, reflectance values corresponding to the points
where there are available depth measurements (reference points)
and where depth simulations will be made (key points) are
extracted from input multispectral images. The key points are so
selected that they do not have an individual behavior and are
evenly distributed throughout considered region. To get model
inputs, the input reflectance on the reference and key points and
depth measurements (expected outputs) on the reference points
are scaled so as to always fall within a specified range that is usable
with the transfer function (see Eq. (2)). The scaled data of reference
points are separated into two parts as training and testing data set.
The scaled data of the key points are used in depth simulations.
Once the model inputs are prepared, model architecture and
parameters can be readily determined using principles in the
Section 3.1. For determined network architecture, model training
and testing are conducted until an acceptable error is reached. The
eventual ANN model tested is used for depth simulations on
selected key and reference (if required) points. The depth estimations are obtained by rescaling network outputs. These estimates
are used to interpolate a contour map or to generate a digital
elevation model.
4. Application and results
Single Band Algorithm (SBA) and Principal Component Algorithm (PCA) did not provide reliable depth estimations in the study
area since the bottom vegetation highly affected the reflectance
from water column. Table 1 shows the coefficients of determination
R2 of the SBA and PCA models. Where R2 designates the variance
ratio of in-situ depth measurements explained by the SBA and PCA
models. SBA and PCA models set for Aster images provided more
reliable result than those for the Quickbird images. SBA models set
for the first and third band Aster images are more reliable than the
model set for second band Aster image, and also the PCA model set
for the first PC is the only efficient model. The SBA model set for the
Quickbird’s third band image seems preferable among the models
set for the Quickbird images. The PCA model set for the first PC of
the Quickbird images is the only PCA model having high R2. In
overall evaluation, though some of the SBA and PCA models may be
used for a course evaluation of the bathymetric structure of the
region, they are not adequate for a reliable bathymetric mapping
because they are not able to explain the variance of the depth
variations sufficiently (see Table 1).
The ANN methodology was applied apropos to flow chart in
Fig. 4. Boolean filter was used, for all band images, to discard the
pixels representing the land area, clouds and vessels and thus to
obtain images only from the marine environment. The Gaussian
Ö. Ceyhun, A. Yalçın / Estuarine, Coastal and Shelf Science 89 (2010) 89e96
93
Image acquisition
(Band images)
In-situ depth measurements
Coordinate
transformation
Image filtering
Extraction of reflectance on the
reference points for each band
Selection of the key points
on which depth estimations
will be made; Extraction of
reflectance on these points
Scaling
Scaling
Testing data
(Scaled reflectance and in-situ
depth measurements)
Training data
(Scaled reflectance and in-situ
depth measurements)
Selection of the
ANN structure
ANN outputs
and
output weights
Training of
the ANN
Initial
weights
and biases
Testing of
the ANN
ANN outputs
and output weights
Re-training
Decision for expectable error between
expected and calculated outputs
(Is the RMSE acceptable?)
No
Yes
Simulation
(ANN outputs on the key/reference points)
Rescaling
Contouring and Digital Elevation Map
Fig. 4. Depth estimation using remote sensed images and ANN.
Table 1
The coefficients of determination for the SBA and PCA models.
Aster
Quickbird
Single band
algorithm
Principal
component
algorithm
Single band
algorithm
Principal
component
algorithm
Band
number
R2
P.C. No
R2
Band
number
R2
P.C. No.
R2
1
2
3
0.75
0.67
0.76
1
2
3
0.74
0.02
0.02
1
2
3
4
0.29
0.58
0.66
0.51
1
2
3
4
0.66
0.01
0.02
0.01
5 5 filter for the Aster images and 7 7 filter for the Quickbird
images were used to reduce scattering effects, noise, and problems
caused by the pixels having individual reflective properties. After
the filtered images and depth measurements were transformed
into same reference system (UTM 35 North, WGS 84), 300 key
points for the Aster model and 450 key points for the Quickbird
model were specified for simulations. 300 reference points were
considered in the analyses. The reflectance values on the reference/
key points were extracted for all band images via a computer
program written in the Visual Basic environment. Depth
measurements on reference points and corresponding reflectance
values were arranged into two (training and testing) data sets.
94
Ö. Ceyhun, A. Yalçın / Estuarine, Coastal and Shelf Science 89 (2010) 89e96
These data sets include input vectors consist of reflectance in
considered spectral bands and expected output vector consist of insitu depth measurements. Reflectance values on the key points for
different spectral bands were arranged into input vectors of
simulation data All input and expected output vectors constituted
were scaled by 0.8 (Rp Rmin)/(Rmax Rmin) þ 0.1 to make them
usable for the ANN analyses. Where Rp designates pth observation
of the input or output vectors; Rmin and Rmax are the minimum and
maximum observations in the vector considered, respectively.
Two ANN models, one for the Aster image inputs and the other
for Quickbird image inputs were set. These models will be called
hereafter as Aster and Quickbird ANN models, respectively. The
architectures of the Neural Networks were determined so that the
numbers of nodes and layers as well as the RMSE between calculated and measured depths would be a minimum. Two inner layers,
each of which has four nodes, were used in the eventual ANN
models. Since only the Aster first three band images and Quickbird
first four band images have significant difference in their reflectance values on the sea surface, three and four neurons were used
in the input layers respectively for Aster and Quickbird models.
LevenbergeMarquardt algorithm was used to train the networks
constituted. Initial weights were given randomly, and log sigmoid
function was used as a transfer function. The significant epoch
number, the epoch after which RMSE values begin to be level off
was used to avoid over-training errors. It was calculated as 3 and 5
for Aster and Quickbird ANN models, respectively. The models
trained were run for the testing data set. Regarding the Aster and
Quickbird ANN models, the depth estimations obtained for the
testing and training data sets were depicted against to the depth
measurements (Fig. 5).
a
According to the results in Fig. 5(a) and (b), the calculated variance ratios explained by ANN models were accepted satisfactory.
Therefore, constituted ANN models were used in depth simulations,
taking the simulation data set as model input. The distanceweighted averages of the simulated depths were used to produce
bathymetric maps. Resultant maps were provided in 5 m precision
to facilitate comparison. Fig. 6 shows the bathymetric maps derived
for the Aster and Quickbird images.
Finally, different band image combinations were used in
modeling to evaluate the effects of the different bands on model
performances. In these analyses, simulations were made for 15
epoch. Resultant depth estimations were compared with in-situ
depth measurements. Table 2 shows the R2 and RMSE between
estimated and measured depths for each band combination used.
5. Discussion
For many coastal engineering and coastal science applications, it is very important to know bathymetry of study area.
However, it is generally very difficult to get detailed information
for reliable bathymetric mapping. Implementing bathymetric
survey for whole study area is more expensive and time/labor
consuming process. The proposed methodology presents a practical solution to estimate water depths in shallow waters via
Artificial Neural Networks (ANN), using remote sensing images
and sample depth measurements. Nonlinear, model free structure of ANN enables the proposed methodology to produce more
accurate estimations than classical regressive methods. Areal and
spatial flexibility of remote sensing images coupled with
nonlinear estimation capability of ANN account for the proposed
b
y = 0.9004x + 1.7331
R2=0.9004
Testing data set
Estimated Depths (m)
Estimated Depths (m)
Training data set
45
40
35
30
25
20
15
10
5
0
45
40
35
30
25
20
15
10
5
0
0 5 10 15 20 25 30 35 40 45
0 5 10 15 20 25 30 35 40 45
Measured depths (m)
Measured depths (m)
y = 0.9187x + 0.0259
R2=0.9184
Testing data set
Estimated Depths (m)
Estimated Depths (m)
Training data set
45
40
35
30
25
20
15
10
5
0
y = 0.8968x + 2.1102
R2=0.8997
45
40
35
30
25
20
15
10
5
0
y = 0.9198x + 0.0172
R2=0.9143
0 5 10 15 20 25 30 35 40 45
0 5 10 15 20 25 30 35 40 45
Measured depths (m)
Measured depths (m)
Fig. 5. Measured vs. estimated water depths, (a) Aster, (b) Quickbird ANN models.
Ö. Ceyhun, A. Yalçın / Estuarine, Coastal and Shelf Science 89 (2010) 89e96
95
Fig. 6. Bathymetric maps derived for (a) Aster, (b) Quickbird ANN models.
methodology’s prevailing advantages especially for preliminary
coastal engineering studies as well as for practical scientific
applications. Main advantages of the proposed methodology can
be listed as below.
provides depth estimations using raw reflectance values,
without refining scattering caused by environmental factors
such as bottom material and vegetation.
allows considering nonlinear multi-parameter relationship
between reflectance from different spectral bands and water
depths.
reduces spatial and repetitive depth measurement requirement.
provides fast and practical bathymetric information.
For the analyses implemented for the study area, SBA and PCA
models produced low accurate depth estimations for Aster images
This may be explained by the fact that since Foca bay is subject to
extensive bottom vegetation, the reflectance from water column is
mainly composed of the reflectance sourced from depth changes and
the reflectance sourced from bottom material. Aster and Quickbird
ANN models provided more efficient estimations. Aster images have
lower resolution e 15 m 15 m e and fewer pixels therefore there is
no possibility to estimates water depths for regions smaller than
225 m2. However, when precise depth measurements can not be
provided, Aster images can be conveniently used by regarding that
reflectance of Aster’s large pixels may produce average estimations.
Aster sensor’s digitations is limited by 256 values, therefore it may
not account for high reliable depth estimations especially for the
regions having dense bottom vegetation and turbidity. It should be
noted that 256 is the full resolution of Aster’s sensors, the range on
the sea surface is much lower than those on the land area and
changes depending on the physical and chemical properties of water
column. The cleaner water will result in larger range and thus in
Table 2
Effects of band numbers on depth estimation performances.
Aster
Quickbird
2
Bands
RMSE
R2
4,3,2,1
3,2,1
2,1
1
0.005
0.007
0.011
0.021
0.92
0.90
0.83
0.67
Bands
RMSE
R
3,2,1
2,1
1
0.006
0.008
0.010
0.90
0.85
0.82
accurate estimations (Leu and Chang, 2005; Louchard et al., 2003).
Quickbird pan-sharpened images have 0.61 m 0.61 m pixel size.
Even local depth changes can be theoretically estimated. However,
since the number of the pixels in a study area is larger for the
Quicbird images, training and testing data sets should be larger to
get reliable models, this means that more in-situ measurement is
needed for the models set for Quickbird images. Correct data couples
may not be easily provided for the Quickbird models since the pixel
size is too small to match depth measurements to the reflectance
values. On the other hand, 11 bit digitation of the Quickbird images
make possible to evaluate small changes on reflective properties of
the water column; therefore more precise estimations can be made
by the Quickbird ANN models.
The number of depth measurements affects remarkably the
accuracy of the bathymetric mapping. This effect can be easily seen
visually comparing the maps generated using different numbers of
in-situ depth measurements (Fig. 2). Different depth values especially near the coast line and on the deep waters are clear. Some zero
depth regions appear both on Fig. 2(a) and (b). The area of such zero
depth regions is lower for Fig. 2(b) than for Fig. 2(a). There are some
local depth regions on Fig. 2(b). These regions might be generated as
only numerical products during map generation. Mentioned effects
of the number of depth measurements on bathymetric mapping
could be reduced by means of the proposed methodology. Bathymetric maps derived for the Aster and Quickbird ANN models (Fig. 6)
presented more accurate and detailed bathymetric information than
generated maps (i.e. Fig. 2). Due to the number and structure of the
data used in bathymetric mapping, the map derived for the Quickbird image inputs (i.e. the map in Fig. 6(b)) has higher precision than
the map derived for the Aster image inputs (i.e. the map in Fig. 6(a)).
The area of zero depth regions is much lower for Fig. 6(b) than for
Fig. 6(a). Different depth regions are clear in Fig. 6(a). The local depth
regions appearing in Fig. 2(b) seems to be smoothed in Fig. 6(b). Note
that reliabilities of estimations on these local regions depend highly
on sampling data used in training and testing. It is seen from the
Fig. 5 that depth estimations lose their effectiveness for the higher
depths than 40 m since the reflectance on sea surface does not
change significantly for deeper regions. This depth value can be
different for other study areas, regarding water clarity.
To reveal efficiency of ANN based bathymetric mapping, the
derived maps (i.e. Fig. 6(a) and (b)) and generated maps (i.e. Fig. 2
(a) and (b)) may be differenced. In-situ measurements are point
96
Ö. Ceyhun, A. Yalçın / Estuarine, Coastal and Shelf Science 89 (2010) 89e96
measurements contrary to the proposed methodology estimating
an average depth value corresponding to the mean reflectance of
the considered pixel. Unless there is no available high resolution
bathymetric map for the study area, differencing may not be
informative especially when images with large pixel are used in
depth estimations. In this study, investigation of the effectiveness
of the proposed methodology is confined with the evaluation of the
training/testing results and the visual comparison of the derived
and generated maps since there is no available high resolution
digital bathymetric map for the region.
In the practice, bottom material, vegetation or pollution affect
the reflectance from the water column, remarkably. Therefore, the
input data should be representative and concurrent to build a reliable model. It is natural that the more changing properties of water
column, the less reliable depth estimates; however, reliabilities of
the model can be increased using large, representative data sets.
Although, theoretically, simultaneous depth measurements and
image acquisition are needed for modeling, data from different
dated images and depth measurements may also be used under the
assumption that bathymetric changes are small or homogeneous on
all reference points during the date difference. However, accuracies
of the results will be depended on how much depth chances have
occurred on reference points between the dates of image acquisition and depth measurements. Accordingly, temporal flexibility in
bathymetric mapping can be provided when depth measurements
from locations where there is no significant bottom current
changing bathymetry are used in modeling. Short term local
pollutants will affect the model performance negatively if reference
points used in training and testing and key points used in simulations are close to or under the effect of the source of pollutants.
Model performance is also affected by the number of input bands
used. The infrared band has no contribution for Aster model and
little contribution for Quickbird model. Water depths may be estimated with moderate accuracy even using only first two bands.
6. Conclusion
Bathymetric map generation process highly depends on the
number of in-situ depth measurements. Depth models provide
practical solutions to increase reliability of generated bathymetric
maps. However, changing bottom vegetation and material makes
difficult to use classical, regressive, optic remote sensing based
depth estimation methods such as SBA and PCA. The methodology
proposed herein provides a useful tool for bathymetric mapping for
practical purposes. Its nonlinear model free structure allows
considering nonlinear relationships between multi-band image
reflectance and water depths, thus more accurate depth estimations
could be obtained. Since the ANN models e a black box model e do
not have to consider the factors affecting the reflective properties of
water column and do not have to necessarily be trained by newer insitu measurements, the methodology can be practically applied
without handling any complex reflectance separation process, and it
can be reliably used for repetitive bathymetric mapping provided
that there is available representative input data set.
Since the study area is subject to intensive bottom vegetation,
SBA and PCA models did not provide reliable depth estimations.
The proposed methodology produced quite accurate estimates.
Even first two bands of both satellite images could model the
bathymetry by a determination coefficient R2 over 80%. Low digitation and spatial resolution of the Aster images facilitated data
manipulation and map generation procedure. The model set for the
Aster images did not provide as much reliable estimates as the
model set for the Quickbird images and could not reveal precise and
local depth changes; however it can be used properly when there is
no need for detailed mapping. On the other hand, high resolution of
the Quickbird images permitted to evaluate local depth changes.
High digitation of the quickbird images made possible more precise
depth estimations.
It is concluded in the study that the Artificial Neural Networks
can be effectively used for bathymetric modeling. The proposed
methodology can produce efficient depth estimations up to
40e45 m depending on water clarity and regardless the portion of
the reflectance sourced from environmental factors. Time dependency of the bathymetric mapping may be decreased by training
ANN models with newer images. Thus, the cost, labor and time
spend for detailed and repeated depth measurements could be
significantly reduced.
Acknowledgements
We would like to thank Dr Hüsnü ERONAT from Dokuz Eylul
University The Institute of Marine Sciences and Technology for his
kind helps on data provision.
References
Atkinson, P.M., Tatnall, A.R.L., 1997. Neural Networks in remote sensing. International Journal of Remote Sensing 18 (4), 699e709.
Civco, D.L., 1993. Artificial neural networks for land cover classification and mapping.
International Journal of Geographic Information Systems 7 (2), 173e186.
Fonstad, M.A., Marcus, W.A., 2005. Remote sensing of stream depths with hydraulically assisted bathymetry (HAB) models. Geomorphology 72 (4), 320e339.
Freeman, J.A., Skapura, D.M., 1991. Neural Networks Algorithms, Applications and
Programming Techniques. Addison-Wesley Publishing Company Inc., ISBN
0-201-51376-5.
Grilli, S.T., 1998. Depth inversion in shallow water based on nonlinear properties of
shoaling periodic waves. Coastal Engineering 35 (3), 185e209.
Hagan, M.T., Demuth, H.B., Beale, M.H., 1996. Neural Network Design. PWS
Publishing, Boston.
Hagan, M.T., Menhaj, M., 1994. Training feed forward networks with the Marquardt
algorithm. IEEE Transactions on Neural Networks 5 (6), 989e993.
Härmä, P., Vepsäläinen, J., Hannonen, T., Pyhälahti, T., Kämäri, J., Kallio, K.,
Eloheimo, K., Koponen, S., 2001. Detection of water quality using simulated
satellite data and semi-empirical algorithms in Finland. The Science of the Total
Environmental 268, 107e121.
Haykin, S., 1994. Neural Networks: A Comprehensive Foundation. MacMillan
College Publishing Co., New York.
Hengel, W.V., Spitzer, D., 1991. Multi-temporal water depth mapping by means of
Landsat TM. International Journal of Remote Sensing 12 (4), 703e712.
Jain, A.K., Moa, J., Mohiuddin, K.M., 1996. Artificial Neural Networks. Theme Feature,
pp. 31e44.
Jain, A.K., 1989. Fundamentals of Digital Image Processing. Prentice Hall, New Jersey.
Leu, L., Chang, H., 2005. Remotely sensing in detecting the water depths and bed
load of shallow waters and their changes. Ocean Engineering 32, 1174e1198.
Lek, S., Delacoste, M., Baran, P., Dimopolos, I., Lauga, J., Aulanier, S., 1996. Application
of Neural Networks to modeling nonlinear relationships in ecology. Ecological
Modeling 90, 39e52.
Louchard, E.M., Reid, R.P., Stephens, F.C., Davis, C.O., Leathers, R.A., Downes, T.V.,
2003. Optical remote sensing of benthic habitats and bathymetry in coastal
environments at Lee stocking Island, Bahamas: a comparative spectral classification approach. Limnology and Oceanography 48 (1, 2), 511e521.
Lyzenga, D.R., 1985. Shallow-water bathymetry using combined lidar and passive
multispectral scanner data. International Journal of Remote Sensing 6 (1),15e125.
Lyzenga, D.R., 1981. Remote sensing of bottom reflectance and water attenuation
parameters in shallow waters using aircraft and Landsat data. International
Journal of Remote Sensing 2, 71e82.
Lyzenga, D.R., 1978. Passive remote sensing techniques for mapping water depth
and bottom features. Applied Optics 17, 379e383.
Lyon, J.G., Lunetta, R.S., Williams, D.C., 1992. Airborne multi-spectral scanner data
for evaluating bottom sediment types and water depths of St. Mary’s River
Michigan. Photogrammetric Engineering & Remote Sensing 58 (7), 951e956.
Martin, K.S., 1993. Applications in Coastal Zone, Research and Management, vol. 3.
United Nations Institute for Training and Research, Geneva.
Mas, J.F., 2004. Mapping land use/cover in a tropical coastal area using satellite
sensor data, GIS and Artificial Neural Networks. Estuarine, Coastal and Shelf
Science 59, 219e230.
Spiter, D., Dirks, R.W.J., 1987. Bottom influence on the reflectance of the sea.
International Journal of Remote Sensing 8, 279e290.
Zanial, A.J.M., 1994. New technique for enhancing the detection and classification of
shallow marine habitats. Marine Technology Journal 28, 68e74.